CompressiveVision
Rebuilds images from compressed measurements with WGAN-GP, CBAM attention and DnCNN refinement, and encrypts the measurements over a simulated channel.
The problem
Compressive sensing captures far fewer measurements than pixels. The hard part is rebuilding a clean image from them, and keeping the measurements private while they travel.
How it works
- Images are split into 16×16 blocks and sampled at a 50% measurement rate.
- A WGAN-GP generator with CBAM attention rebuilds the image, and a DnCNN pass removes leftover noise.
- The measurements are encrypted over a simulated transmission channel.
Results
The full pipeline reaches a PSNR of 24.17 dB and an SSIM of 0.765, up from 23.38 dB and 0.739 for the plain WGAN-GP baseline. The tile shows an original image beside its reconstruction.